Two audiences on the same streaming plan needed frequencies six to eight times apart, and for four months the reporting layer showed one tidy number that was correct for neither.
What actually happened
Adduro, a Denver-based managed streaming shop, has published a retail case study that a DMWF sponsored newsletter picked up in late July. The campaign: a national direct-to-consumer retailer selling party and event supplies, running roughly $350,000 across streaming TV and display from February to May 2026, bought direct from Disney+, Hulu, Paramount+, Peacock, Roku, Tubi, Sling TV and VIZIO with no intermediaries in the path.
The plan carried two audiences. Parents buying for birthdays, holidays and classroom parties. B2B buyers stocking schools and events. Both sat under one weekly frequency ceiling. When the delivery data was broken out by segment, parents converted most efficiently at about one CTV impression a week. B2B buyers stayed efficient at six to eight. The agency capped each audience separately, assigned distinct roles to CTV and display, and reconciled platform reporting against point-of-sale revenue every week. Streaming RoAS is reported up 48%, peaking within 30 days of the reset.
The obvious reading is that frequency capping works and someone finally did it properly. That reading is too small. The interesting part is not that the cap was wrong. It is that a 6-to-8x divergence in optimal exposure survived four months and a third of a million dollars without appearing anywhere in the reporting — and that the fix had less to do with the cap itself than with where in the stack it gets enforced.
Why the average swallows the signal
Frequency reported as a mean across a flight is a summary statistic laid over a violently skewed distribution. Delivery to households is long-tailed. Several large-scale analyses put the concentration starkly: only 8% to 15% of households exceed their intended cap, but those households absorb between 42% and 60% of total impressions. Adduro’s own February commentary puts over-exposed delivery at 15% to 25% of impressions. Either way, the mean sits nowhere near where the money actually lands.
Layer a second error on top. Even a correctly measured household frequency is the wrong input if it pools segments with different response curves. Averaging a curve that saturates at one impression with a curve that keeps returning through eight produces a number that is optimal for neither and defensible to nobody. This is the failure mode that scales with budget: the more you spend, the smoother the aggregate looks, which is why the largest advertisers are structurally the least likely to catch it.
Then the enforcement question. A cap set in most platforms is a counter that reports after delivery. Suppression has to happen at the bid, against a household identity resolved across every supply path, or clustering simply reappears inside pods and evenings. Worth noting: the DMWF write-up attributes bid-level frequency governance to GammaBurst. Adduro’s own product documentation puts pre-bid household frequency coordination in GammaRay, its in-house bidder; GammaBurst is the reporting and analytics layer. The distinction is exactly the point the article was making.
Generalising the mechanism:
- Never optimise against a mean drawn from a skewed distribution. Demand the distribution by frequency band.
- Segment before you cap. Response curves belong to buyer types, not to campaigns.
- Enforcement beats reporting. A ceiling counted after the fact is a description, not a control.
- Pooled metrics degrade as spend rises. Aggregation is a function that gets more flattering the more data you feed it.
Where the number gets soft
The window and the peak do not match. The lift peaked within 30 days of the reset on a campaign that ran roughly sixteen weeks. A peak is the best month, not the sustained level. The sustained figure is not disclosed, and the DMWF version blurs the two by placing “30 days” and “16 weeks” in the same sentence.
Seasonality is doing unquantified work. Party and event supplies, February through May, covers Easter, proms, end-of-year school events and graduation. B2B buying for schools ramps hard into precisely that window. The reset raised B2B exposure at the moment B2B demand was rising on its own. There is no holdout and no geo control to separate the two.
Reconciling to the register is better, not causal. Matching platform numbers to point-of-sale revenue strips out platform self-report bias, which is a real improvement over a dashboard figure. It does not strip out baseline. Practitioners running holdout programmes routinely find platform-attributed RoAS overstating incremental RoAS by 20% to 60%. The IAB CTV measurement guidance the case study cites standardises counting, not causation.
Three variables moved at once. Caps split, channel roles split, reporting cadence changed. The 48% cannot be assigned to frequency governance alone, and the case study does not claim it was isolated.
Only one channel was actually personalised. The detail the aggregator dropped: display ran at 10 to 25 impressions a week for both audiences, held steady, while CTV split apart. The story of two very different buyers applies to one channel. The other stayed averaged.
The premise itself is contested. An ANA and Innovid study found average CTV frequency of 4.6 with 85% of households seeing an ad once or twice, and a separate analysis across 20 major advertisers concluded the CTV frequency problem was highly exaggerated — with the real culprit being duplicated supply sold by multiple providers rather than careless capping.
What the move signals
Frequency is drifting from a settings problem to a supply-chain and identity problem, and the two available answers are opposites.
The walled-garden answer removes fragmentation rather than measuring it. Amazon DSP absorbing Netflix inventory puts long-form streaming, live sports, AVOD and retail media behind one household graph, which makes deduplication a property of the platform rather than a service you buy. The independent answer runs the other way: buy direct from publishers, own the bidder, resolve identity against third-party graphs. Adduro is executing that second play while also carrying Amazon Ads verified partner status, which tells you how firmly the middle market has committed to either side.
The defensible asset in this market is not the cap. It is the household graph plus the ability to act on it before the bid clears. Reporting layers are commoditising fast; every vendor now ships a frequency distribution chart. The scarce thing is suppression authority across paths you do not own.
What to take into your own roadmap
- Pull the frequency distribution by band for your last completed flight. Ask specifically for the share of impressions delivered to households already above your intended ceiling.
- Model your response curve separately for each buyer type before you touch a single cap. If you cannot name the segments, you cannot set the ceiling.
- Ask your DSP in writing whether the cap suppresses pre-bid or counts post-delivery. Treat a vague answer as a no.
- Map every path into a household — each DSP, each direct buy, owned-and-operated, linear extensions — and count paths rather than campaigns.
- Run a geo or audience holdout alongside the change. Two weeks of withheld markets converts a correlation into a figure your finance lead will accept.
- Change one variable per test. Splitting caps, reassigning channel roles and altering reporting cadence simultaneously buys you a result you cannot reproduce.
- Keep the weekly reconciliation running after the win. Peaks are cheap to generate; the sustained level is the only number worth budgeting against.
The sequence matters more than the tactic. Measure the distribution, segment the curve, enforce at the bid, then prove causality with a holdout. Run that order backwards and you end up with a 48% that is real to the people who produced it and unreproducible to everyone else.
The trade-off nobody advertises: pre-bid suppression costs reach, and it requires an identity graph you almost certainly do not own. The strategic decision in front of you is not whether to cap by audience. It is whose graph you are willing to rent, and what that dependency costs when the rent goes up.
